PK Vishwanathan, Professor at Great Lakes Institute of Management; speaks at Cypher 2017
summary
Argues that neural networks are not a complete black box and demonstrates a practical way to extract relative importance by running sensitivity analysis on inputs rather than trying to interpret raw weights. Two cases: a 30‑case marketing set where neural net classifies 27/30 (90%) vs logistic 24/30 (84%), and a 1,000‑record mailing set where overall accuracy is 92.4% vs 92.2% but neural net correctly identifies 14/83 buyers vs 8/83 for logistic. Notes lack of formal hypothesis tests and unstable weights due to local optima.
AI-written summary · about
chapters (4)
related talks
48mKeynote – Sameer Singh
25mWhy No Model Is A Black Box
27mData Science and Neural Networks - Big Love or just a Love Affair?
37mMachine Learning, The Right Way
32mMLDS 2019: Pradeep Gulipalli of Tiger Analytics at Machine Learning Developers Summit 2019
06mNeural Networks Demystified: Part 7
39mKeynote 3: How did you know? Explaining Black Box Model Predictions in Python
31mWait, You Test Neural Networks Too?